Spaces:
Running on Zero
Running on Zero
File size: 13,474 Bytes
92d53f7 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 | """The inference entry point. Conversation in, calibrated tier out.
python -m lighthouse.model.predict --fit # fit on all 80 rows, write the artefact
python -m lighthouse.model.predict # show predictions for the corpus
## Why this file exists at all
`conversation_head.py` answers "how good is this model", and it answers it honestly:
everything there is cross-validated, and `main()` fits the full model exactly once, prints
the coefficient table, and throws it away. That is correct for a report and useless for a
product. Nothing persisted a fitted estimator, nothing persisted the isotonic calibrator,
and there was no `predict(conversation) -> Tier` anywhere in the repo β so days 6 and 7,
which both assume an escalation card exists, had nothing to build on.
This is the missing half: fit once on everything, write the parameters down, and expose a
function that turns one conversation into one tier.
## Weights as JSON, not as a pickle
The artefact is `data/artifacts/conversation_head.params.json`: scaler mean and scale,
logistic coefficients and intercepts, and the isotonic calibrator's knots. Three reasons,
in order of how much they matter:
1. **A pickle is executable.** Loading one runs arbitrary code from the file. This model
decides whether a child's case reaches a counsellor; the artefact that encodes it
should be inspectable text, not a payload.
2. **It survives version drift.** A joblib pickle is bound to the scikit-learn version
that wrote it, and day 9 deploys to a Space we have not built yet.
3. **It is portable.** Standardise, dot product, softmax, interpolate β a linear model is
arithmetic. If the day 9 Space is more trouble than it is worth, the same numbers run
in TypeScript with no Python at all. That option stays open because of this format.
## What this deliberately does NOT do
It does not apply the gate. `predict_tier` returns what the *model* thinks, and the caller
passes it through `gate.safety.apply_verdict` to get what the *product* says. Keeping the
floor out of this file is what keeps "no model output may lower a gate floor" enforceable
in one place. `predict_case` below is the composed version, and it is the one callers want.
"""
from __future__ import annotations
import argparse
import json
from dataclasses import dataclass
from typing import Sequence
import numpy as np
from sklearn.isotonic import IsotonicRegression
from lighthouse import config
from lighthouse.data.synthetic import load, load_turn_probs
from lighthouse.gate.safety import SafetyVerdict, apply_verdict, evaluate_conversation
from lighthouse.model.conversation_head import ESCALATED, _model, _ranker, build_rows
from lighthouse.model.features import ABLATIONS, featurise, names_for
from lighthouse.taxonomy import TIER_ORDER, Tier
PARAMS_PATH = config.ARTIFACTS_DIR / "conversation_head.params.json"
#: The feature set the shipped model uses. `full` rather than `+report` even though
#: `history` contributes exactly zero on this corpus (day 4), because day 6 seeds returning
#: students and the column has to already be there when it stops being inert.
SHIPPED_GROUPS = ABLATIONS["full"]
# --------------------------------------------------------------------------------------
# Fitting
# --------------------------------------------------------------------------------------
def _dump_pipeline(pipeline) -> dict:
"""Pull a `StandardScaler -> LogisticRegression` pipeline apart into plain numbers."""
scaler, clf = pipeline[0], pipeline[-1]
return {
"scaler_mean": scaler.mean_.tolist(),
"scaler_scale": scaler.scale_.tolist(),
"coef": clf.coef_.tolist(),
"intercept": clf.intercept_.tolist(),
"classes": [int(c) for c in clf.classes_],
}
def fit_and_save() -> dict:
"""Fit the tier head, the escalation ranker, and the calibrator on **all** 80 rows.
Fitting on everything is correct here and would be wrong in `conversation_head.py`.
That file measures generalisation, so it never lets a model see the row it is scored
on. This file ships a model, and a model shipped to a school should have been trained
on every example we have. The honest performance numbers stay where they were measured:
cross-validated, in `docs/results.md`. Do not quote a number produced by this file.
"""
rows = build_rows()
X = np.vstack([r.features.vector(SHIPPED_GROUPS) for r in rows])
y = np.array([r.tier.rank for r in rows])
needs_counsellor = np.isin(y, [t.rank for t in ESCALATED]).astype(int)
head = _model()
head.fit(X, y)
ranker = _ranker()
ranker.fit(X, needs_counsellor)
# The calibrator is fitted on out-of-fold ranker scores, never on the in-sample ones.
# In-sample scores from a model that has seen every row are near-separable, and an
# isotonic fitted on those collapses to a step function that reports 0.02 or 0.98 and
# nothing in between β perfectly confident and completely uninformative.
from lighthouse.model.conversation_head import _inner_escalation
inner_raw, inner_true = _inner_escalation(X, needs_counsellor)
iso = IsotonicRegression(out_of_bounds="clip", y_min=0.0, y_max=1.0)
iso.fit(inner_raw, inner_true)
params = {
"note": (
"Generated by `python -m lighthouse.model.predict --fit`. Fitted on all 80 "
"synthetic conversations. These weights are for INFERENCE ONLY β every "
"performance number in docs/results.md is cross-validated and comes from "
"conversation_head.py, not from a model fitted on its own test set."
),
"seed": config.SEED,
"n_train": len(rows),
"feature_groups": list(SHIPPED_GROUPS),
"feature_names": names_for(SHIPPED_GROUPS),
"tier_order": [t.value for t in TIER_ORDER],
"head": _dump_pipeline(head),
"ranker": _dump_pipeline(ranker),
"isotonic": {
"x": iso.X_thresholds_.tolist(),
"y": iso.y_thresholds_.tolist(),
},
}
config.ARTIFACTS_DIR.mkdir(parents=True, exist_ok=True)
PARAMS_PATH.write_text(json.dumps(params, indent=2) + "\n")
return params
# --------------------------------------------------------------------------------------
# Loading and predicting
# --------------------------------------------------------------------------------------
@dataclass(frozen=True)
class TierPrediction:
"""What the model alone says. Not what the product says β see `predict_case`."""
tier: Tier
confidence: float
"""P(tier), from the five-class head."""
escalation: float
"""Calibrated P(needs a counsellor). This is what the queue is ranked by."""
proba: dict[str, float]
@dataclass(frozen=True)
class CasePrediction:
"""What the product says: the model, constrained by the gate."""
tier: Tier
"""Post-gate. This is the tier on the escalation card."""
model_tier: Tier
"""Pre-gate, kept so a card can show that the gate moved it."""
confidence: float
escalation: float
proba: dict[str, float]
verdict: SafetyVerdict
floor_reason: str | None
"""Non-null when the gate raised or capped the tier. Rendered verbatim on the card."""
@property
def queue_rank(self) -> float:
"""Sort key for the counsellor queue: `floor_rank + escalation`.
Day 4 established this is the product's ranking, not the model's, and that ranking
by the model score alone understated recall@20 by 0.15. The floor is the primary
sort and the model is the tie-break, which is exactly the documented precedence.
"""
floor_rank = self.verdict.floor.rank if self.verdict.floor else 0
return float(floor_rank) + float(np.clip(self.escalation, 0.0, 1.0))
class ConversationHead:
"""A fitted head loaded from JSON. Construct via `ConversationHead.load()`."""
def __init__(self, params: dict) -> None:
self.params = params
self.feature_groups = tuple(params["feature_groups"])
self.feature_names = list(params["feature_names"])
self._head = params["head"]
self._ranker = params["ranker"]
self._iso_x = np.array(params["isotonic"]["x"], dtype=np.float64)
self._iso_y = np.array(params["isotonic"]["y"], dtype=np.float64)
@classmethod
def load(cls, path=PARAMS_PATH) -> "ConversationHead":
if not path.exists():
raise FileNotFoundError(
f"{path} is missing. Run: python -m lighthouse.model.predict --fit"
)
return cls(json.loads(path.read_text()))
@staticmethod
def _decision(block: dict, x: np.ndarray) -> np.ndarray:
z = (x - np.array(block["scaler_mean"])) / np.array(block["scaler_scale"])
return np.array(block["coef"]) @ z + np.array(block["intercept"])
def _softmax(self, block: dict, x: np.ndarray) -> dict[int, float]:
scores = self._decision(block, x)
# Binary logistic regression stores one row of coefficients, not two.
if scores.shape[0] == 1:
p1 = 1.0 / (1.0 + np.exp(-scores[0]))
probs = np.array([1.0 - p1, p1])
else:
shifted = scores - scores.max()
exp = np.exp(shifted)
probs = exp / exp.sum()
return {int(c): float(p) for c, p in zip(block["classes"], probs)}
def predict_tier(self, features) -> TierPrediction:
"""Model only. The gate has not been applied."""
x = features.vector(self.feature_groups)
by_rank = self._softmax(self._head, x)
proba = {t.value: by_rank.get(t.rank, 0.0) for t in TIER_ORDER}
best_rank = max(by_rank, key=lambda r: by_rank[r])
raw = self._softmax(self._ranker, x).get(1, 0.0)
# Piecewise-linear interpolation over the isotonic knots, clipped at the ends β
# the same thing sklearn's `predict` does, without the sklearn dependency.
escalation = float(np.interp(raw, self._iso_x, self._iso_y))
return TierPrediction(
tier=TIER_ORDER[best_rank],
confidence=by_rank[best_rank],
escalation=escalation,
proba=proba,
)
def predict_case(
self,
probs: np.ndarray,
turns: Sequence[str],
*,
verdict: SafetyVerdict | None = None,
prior_sessions: int = 0,
prior_max_tier_rank: int = 0,
) -> CasePrediction:
"""The whole pipeline: features -> head -> gate. What a caller actually wants.
`t4_override` is passed when the head is confidently predicting T4 itself, so a
self-harm phrasing the regex banks never anticipated is not capped into invisibility
by the ceiling. It can only ever raise the outcome, and it is recorded in the reason.
"""
verdict = verdict if verdict is not None else evaluate_conversation(turns)
features = featurise(
np.asarray(probs, dtype=np.float64),
verdict,
turns,
prior_sessions=prior_sessions,
prior_max_tier_rank=prior_max_tier_rank,
)
prediction = self.predict_tier(features)
t4_override = (
prediction.tier is Tier.T4
and prediction.proba[Tier.T4.value] >= config.GATE_HIGH_SCORE
)
tier, reason = apply_verdict(
prediction.tier, verdict, t4_override=t4_override
)
return CasePrediction(
tier=tier,
model_tier=prediction.tier,
confidence=prediction.confidence,
escalation=prediction.escalation,
proba=prediction.proba,
verdict=verdict,
floor_reason=reason,
)
_CACHED: ConversationHead | None = None
def head() -> ConversationHead:
"""Process-wide singleton. Parsing the artefact per request is pure waste."""
global _CACHED
if _CACHED is None:
_CACHED = ConversationHead.load()
return _CACHED
def main() -> None:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--fit", action="store_true", help="refit and write the artefact")
args = parser.parse_args()
if args.fit:
params = fit_and_save()
print(f"wrote {PARAMS_PATH.relative_to(config.REPO_ROOT)}")
print(f" {params['n_train']} conversations, {len(params['feature_names'])} features")
print(f" isotonic knots: {len(params['isotonic']['x'])}")
return
model = head()
probs = load_turn_probs()
conversations = load()
moved = 0
correct = 0
print(f"{'id':<9}{'true':<6}{'model':<7}{'final':<7}{'conf':>6}{'esc':>7} gate")
print("-" * 74)
for convo in conversations:
case = model.predict_case(np.array(probs[convo.id]), convo.student_turns)
if case.tier is not case.model_tier:
moved += 1
if case.tier is convo.tier:
correct += 1
flag = "" if case.tier is case.model_tier else "<- gate"
print(
f"{convo.id:<9}{convo.tier.value:<6}{case.model_tier.value:<7}"
f"{case.tier.value:<7}{case.confidence:>6.2f}{case.escalation:>7.2f} {flag}"
)
print(
f"\n{correct}/{len(conversations)} exact tier match in-sample "
f"(NOT a generalisation estimate β see docs/results.md), "
f"gate moved {moved}"
)
if __name__ == "__main__":
main()
|